| Summary: | 碩士 === 國立臺灣大學 === 電機工程研究所 === 84 === In order to gain an insight into the characteristics of bit-
rate variation, we characterize and model a whole length MPEG
bit-rate sequence in the thesis. We propose a model to grasp
the long-term and short-term correlations of bit-rate
variation. As there are I, P, and B types of frames in an MPEG
video sequence, our model is composed of three subsequences.
The subsequences are self-similar, each with long-range
dependence in itself and short-range dependence with each
other. In looking for the bit-rate distriution, we find that
for P- and B-frames, the whole picture sequence and the `group
of pic- tures' (GOP) sequence, a PT5 density function fits the
distribu- tion better than the frequently used gamma density
function in the literature. The latter is suitable for I-
frames, however. In analyzing the sample autocovariances, we
point out the long-range dependence (LRD) and self-similarities
in an MPEG source and use various methods to estimate the Hurst
parameter. In addition, from the perspective of coding
algorithm, we point out the non-stationarity in an MPEG
sequence and propose appro- priate calculations of time-
averaged sample autocovariances, wherein the separate means of
I-, P- and B-frames are taking into account. In the generation
of synthetic MPEG sequences, a "two-layer" synthesis approachis
adopted. The time-averaged autocovariances are modeled by a
Gaussian sequence, and the marginal distribut- ions of separate
frame types are modeled by transforming the Gaussian sequence
into the corresponding density functions. Via the
transformation, the non-stationarity of the model is simul-
taneously restored. From the statistical results, the model be-
haves well in modeling histograms and LRD, but slightly looses
its accuracy in matching autocovariances at large lags.
|
|---|
